Embedding inference
AssigneePinterest, Inc.
InventorsHeath Vinicombe, Chenyi Li, Yunsong Guo, Yu Liu
Life of this patent
Estimating the term…
Abstract
Systems and methods are presented for inferring an embedding vector of an item of a first type into the embedding space. Upon receiving a first time for which there is no embedding vector, documents of a document corpus that include (co-occurrence) both the received item and other items of the same type are identified. Of those other items that have embedding vectors, those embedding vectors are retrieved and averaged. The resulting averaged embedding vector is established as an inferred embedding vector for the received item.
Claims13
A computer-readable medium bearing computer-executable instructions which, when executed by a processor on a computing system, carry out a method for inferring an embedding vector for a target content item of a target content type, the method comprising:
- accessing document corpus information, the document corpus information describing a plurality of documents that include items of a plurality of content types including content items of the target content type;
- identifying a set of documents of the document corpus that include the target content item according to the accessed document corpus information;
- for each document of the identified set of documents that include the target content item, adding content items of the target content type within a current document to an aggregated set of content items;
- for each content item of the target content type within the aggregated set of content items:
- determining a pointwise mutual information (PMI) score for a current content item in view of the target content item;
- and
- if the PMI score for the current content item fails to at least meet a threshold PMI score for the target content item, removing the current content item from the aggregated set of content items;
- accessing embedding vectors of the content items of the target content type remaining in the aggregated set of content items;
- averaging the accessed embedding vectors;
- and
- associating the resulting averaged embedding vector with the target content item as an inferred embedding vector for the content item.
A computer system for determining a similarity value of a first content item of a first content type to a second content item of a second content type, the system comprising at least a processor and a memory, wherein the processor, in execution of one or more executable components, configures the computer system to:
- determine a first embedding vector for the first content item for a first embedding space of the first content type;
- determine a second embedding vector for the second content item for a second embedding space of the second content type;
- identify a predetermined first number of closest items of the first content type to the first content item, each of the first number of closest items of the first content type having an established association to one or more content types of the second content type;
- for each of the closest items to the first content item, identify a top content item of the second content type associated with a current closest item;
- determine an averaged embedding vector in the second embedding space from embedding vectors of the top content items associated with the closest items to the first content item;
- and
- determine a first similarity score between the first content item and the second content item according to a measure of similarity between the averaged embedding vector in the second embedding space and the second embedding vector of the second content item in the second embedding space.
Dependencies are inferred from the claim text.